# A framework for the emergence and analysis of language in social learning agents

**Authors:** Tobias J. Wieczorek, Tatjana Tchumatchenko, Carlos Wert Carvajal, Maximilian F. Eggl  
**arXiv:** [2305.02632](https://arxiv.org/abs/2305.02632)  
**Date:** 2023-05-04  
**Summary page:** https://www.summarizepaper.com/en/paper/2305.02632v1/

## Key points


## Summary

The use of artificial neural networks (ANNs) as research models has become increasingly popular, but questions remain about their generalizability and representational invariance. In contrast, biological neural networks evolved under social constraints to enable communicable representations, demonstrating impressive generalization capabilities. To bridge this gap between natural and artificial systems, a recent study proposed a communication protocol between cooperative agents to analyze the formation of individual and shared abstractions and their impact on task performance. This communication protocol aims to mimic language features by encoding high-dimensional information through low-dimensional representation. Using grid-world mazes and reinforcement learning, teacher ANNs pass a compressed message to a student ANN for better task completion. Through this process, the student achieves a higher goal-finding rate and generalizes the goal location across task worlds. To further optimize message content and improve information encoding, the study allowed the student to generate messages itself after performing tasks with messages from teachers. These student-generated messages were then passed back to itself for assessment of its performance with these messages. This process created a simple generalist agent that supplied information through the same language encoding that was kept fixed throughout the experiment. However, despite this degradation in message quality over time, RL agent performance was still significantly affected by these degraded messages. Overall, this study highlights the role of language as a common representation between agents and its implications on generalization capabilities. It also suggests that an accurate representation in the space of messages requires bi-directional input which can be achieved by allowing students to generate their own messages after learning from teacher-generated ones.